Anthropic has released FlashProve, an open-source research prototype that uses zkSNARK cryptography to verify that a specific AI model generated a given response without exposing the model weights. The system addresses the trust problem in model routing, where providers could silently substitute cheaper models or attackers could tamper with inference streams. FlashProve works by creating a digital fingerprint of model weights and generating a compact proof per response attesting to correct execution. For Meta's Llama 4 Maverick (400B parameters), proof generation runs at approximately 110 tokens per second on a single H100 GPU, producing 19 MB proofs verifiable in 17 seconds on a standard CPU. This represents a 70x–95x overhead versus standard inference, a dramatic improvement over the prior comparable open-source system, which required over 14 hours for 1,000 tokens, generated a 93.6 GB proof, and needed nearly a day for verification. Anthropic identifies use cases including regulatory compliance verification, safety model audit trails, and server-side tampering detection. The prototype currently supports contexts up to 1,024 tokens, does not conceal model architecture, and only functions with greedy decoding (single most likely token), not the stochastic sampling used in typical chatbot deployments. Code, Lean proofs, and documentation are available on GitHub under Apache 2.0.